# What was tested — 16 September 2026 This is an early local pilot, not a production service or a validated productivity claim. ## Actual Claude Code smoke tests Tested on macOS with Claude Code **2.1.271**, authenticated through the existing Claude subscription. The primary model reported by the CLI was `claude-opus-5` at high reasoning. Isolated fictional projects, fresh nonpersistent sessions, auto-memory and hooks disabled for these controlled tests, no outside source access requested. The harness restricted tools; several shell directory-creation attempts were denied, and the file-write tool completed the local write instead. These restrictions were test settings, not changes to your Claude configuration. Observed and checked: - Save created a readable checkpoint containing the human's stated constraints, with no application edits. - A second save created `000002.md`; the first checkpoint's hash remained unchanged. - A fresh `/bakery-resume` restored the stated no-sign-in, manual-capacity, deferred-waitlist and field-name decisions. - After the project source changed, another fresh resume identified the stale “nothing implemented” statement, read the existing prototype, and proposed working from it rather than making a second form. Source hashes remained unchanged during that read-only return. - Feedback was stored verbatim. A second feedback record preserved the first and did not invent a score or time-saved claim. - A fresh project with no checkpoint reported missing history instead of inventing it. Early testing exposed two issues that were corrected: timestamp filenames could be invented when the clock was unavailable, and a summary equated no sign-in with anonymity. The final candidate uses numbered files and preserves the narrower meaning. Feedback wording also explicitly distinguishes local files from Claude's normal remote processing. The final smoke outputs reflect those corrections. The first live attempts failed because the local OAuth login had expired. They made no inference requests; testing succeeded after the account owner refreshed login. This was an environment issue, not hidden by replacing the record. ## Actual paired demonstration Two fresh Claude sessions received identical fictional PROJECT.md bytes. One had only those source files; the other also had a real generated checkpoint and the b.akery resume instructions. Neither had prior chat or auto-memory loaded. One run per condition; no reruns to obtain a stronger contrast. Both proposed useful work. The source-only response said it could not recover prior conversations or decisions. The checkpoint response explicitly preserved the prior no-sign-in, manual-capacity and deferred-waitlist decisions. Both still lacked actual event details. See [source-only response](examples/compare-control.txt), [checkpoint response](examples/compare-checkpoint.txt) and [checkpoint](examples/checkpoint.md). These are actual outputs; the page's short quotes use ellipses and link to the complete text. This demonstrates retained information in this example. It does not establish overall superiority, a success rate, measured time savings, or an advantage over Claude's native resume or memory. It also does not prove that all statements in a future generated handoff will be accurate. Read the recap while the work is fresh. ## Desktop compatibility Claude’s [Desktop documentation](https://code.claude.com/docs/en/desktop#use-skills) confirms that local Code sessions support project skills. Participants can ask Claude to install the kit from within that conversation, without operating a terminal themselves. James confirmed that the Code tab accepts ZIP attachments; attachment-first setup is now the recommended path. The agent has not independently tested attachment extraction or Desktop UI installation and invocation end to end; live skill tests used the command-line runner. The first trial uses the same local folder with worktree isolation off, because ignored checkpoint files and uncommitted skills do not automatically travel into another worktree or a cloud session. Ordinary Chat and Cowork installation are outside this package. ## Installer and walkthrough The original six installer regression checks passed: normal/repeated installation preserving existing project rules, conflicting-skill refusal before installation, symlink refusal, preservation of existing captures, refusal to replace incompatible ignore rules, and missing-destination handling. A separate Git check confirmed that a new checkpoint is ignored. All three skill files passed format validation; shell and JavaScript syntax checks passed. The three illustrative UI stages, setup dialog, clipboard success feedback and keyboard arrow navigation were exercised in the Codex in-app browser. No console errors were observed. Clipboard transfer to another application was not independently verified. Responsive styling was inspected in the available narrow browser viewport; this is not a comprehensive device/accessibility audit. For v0.1.1, all eleven installer checks passed. Added checks cover worktree-path warnings, linked worktrees outside Desktop’s default path, ZIP/extracted-folder exclusion under an actual `git add -A`, nested-project exclusion scope, and preservation of already-staged files with a warning. Repeated installation preserves existing local exclusion content without duplicating patterns. Skill contents are unchanged from v0.1; no new model runs were needed for this installer/documentation patch. The installer was run on macOS. Linux/WSL and manual installation elsewhere have not been exercised. The optional cross-tool resume prompt has not been independently tested in Cursor or Codex. Model-generated skill behavior is not an operating-system security sandbox. ## Data and costs The distributed kit contains only the starter files and fictional examples. No actual research cases, credentials, private session logs or personal project files are bundled. The kit makes no network requests itself and sends nothing automatically to b.akery; using the skills invokes Claude normally under the user's account. It installs no background jobs or capture hooks. Existing Claude permissions and usage limits apply. CLI token/cost receipts are retained privately in the development workspace. Their price fields are list-price estimates, not reconciled subscription billing. No OpenAI API calls were used for these smoke tests. This release does not implement background comparative testing or automatic transcript capture. Format reference: [Claude Code skills](https://code.claude.com/docs/en/skills).